Forecasting and AI Agents for Arison's Warehouse Operations
Arison tracked inventory on spreadsheets and guessed at demand. We shipped agents for inventory allocation and forecasting, plus a self-service client dashboard — and delivered visibility the team had chased for three years.
Arison operates a multi-site warehouse network for a portfolio of retail and e-commerce clients. Every client expects inventory accuracy, every client calls in for stock checks, and every client pays for the operational overhead that those phone calls generate.
Before this engagement, inventory was tracked on spreadsheets that pulled from four disconnected source systems — ERP, WMS, e-commerce channels, and shipping logs. Demand was forecast by intuition. Stockouts were frequent, inventory cost was the largest line on the working-capital report, and BI attempts had stalled because every dashboard was a day stale the moment it rendered.
sesgo.ai shipped a unified data layer, a forecasting ensemble, AI agents for inventory allocation and reorder decisions, and a self-service dashboard with a clean API boundary — so Arison's clients can now check their own inventory without picking up the phone. The system went live in 12 weeks.
01 · The Challenge
Four systems, zero source of truth, stockouts as a way of life
Arison's inventory reality lived across four systems that did not talk to each other. The ERP held the financial truth, the WMS held the location truth, the e-commerce integrations held the sales truth, and the shipping logs held the outbound truth. Nothing reconciled automatically. Every morning a team pulled exports into spreadsheets and tried to stitch a picture together — a picture that was already obsolete by mid-afternoon.
Demand forecasting was effectively a guess. The planning team used rolling averages and gut feel. That worked well enough for SKUs with stable turnover, but the long tail — seasonal items, promotional bundles, and the newest inventory tier — was where cost hid. The forecasting gap translated directly into two expensive failure modes: over-ordering, which tied up working capital on slow-movers, and under-ordering, which produced frequent stockouts on fast-movers.
The stockout pattern was the most visible problem to clients. Arison was absorbing fourteen stockout events per month across the network. Each event meant missed revenue for the client, a phone call to Arison ops, and an incident report that nobody had time to root-cause before the next one hit.
Finance flagged inventory cost as the largest working-capital drag in the business. The CFO had allocated budget for "something that fixes this" two years running. Two prior BI vendor attempts had stalled — dashboards looked great in demos, but the underlying data was always a day stale, and nobody trusted numbers they could not refresh in the moment.
The client-facing overhead completed the picture. Arison's support team was fielding an average of hundreds of "do I have stock of X?" calls per week from client buyers, operators, and AP teams. That overhead had no margin attached to it, and it was growing every quarter as the client roster grew.
02 · Our Approach
Data layer first, models second, agents last — API-first throughout
The KPI framing made the priority obvious. Working capital is the upstream metric. Stock accuracy, stockout count, and inventory turnover are the leading indicators of working capital and the only three numbers that needed to move. Every decision in the engagement was evaluated against whether it moved those three.
Sequencing followed directly. Unified data first, forecasting second, agents third, dashboard throughout. Nothing downstream could work if the data was wrong upstream, so the first four weeks of the engagement were spent on the unglamorous discipline of ingestion, contracts, and a single canonical inventory model in Snowflake built with dbt.
Forecasting model choice was deliberate. We shipped an ensemble of Prophet and XGBoost rather than the larger neural options. Prophet handled the seasonal signal on mid-volume SKUs where explainability mattered most to the planning team. XGBoost handled the long tail and cross-SKU feature signal. The ensemble was tuned per SKU density tier — different models absolutely win on different segments of the catalog, and assuming one-size-fits-all was the failure mode we most wanted to avoid.
- KPI-first framing Stock accuracy, stockouts, and inventory turnover as leading indicators of working capital. No other metric got airtime until those moved.
- Unified data layer first Four source systems ingested into a single canonical model in Snowflake, built with dbt, contract-tested before any model trained.
- Prophet + XGBoost ensemble Explainable on mid-volume SKUs, robust across the long tail. Per-SKU-density tuning, not one-size-fits-all.
- API-first dashboard Every dashboard metric exposed as a stable API endpoint — so Arison's clients can embed the same signal in their own portals.
The AI agents were the last layer, not the first. The allocation agent consumes forecast output plus current inventory state and calculates reorder points and allocation splits across warehouse locations. The exception-handling agent surfaces anomalies — sudden demand shifts, slow movers that the ensemble flags, supplier lead-time breaches — for human review instead of silently acting. That guardrail-first design is why the CFO accepted the system instead of treating it as a black box.
03 · The Solution
One data spine, one forecast layer, two portals
The system is built around a single canonical inventory model with three downstream surfaces: the forecasting layer, the agent layer, and the portal layer. The portal layer is exposed twice — as an internal operator dashboard for Arison, and as a client-facing self-service portal that clients embed into their own workflows.
SOURCE SYSTEMS CONSUMPTION
+---------------+
| ERP |--+
+---------------+ |
+---------------+ | +--------------+ +---------------+
| WMS |--+--->| Ingestion |--->| Canonical |
+---------------+ | | (Airflow + | | Inventory |
+---------------+ | | dbt) | | Model |
| E-commerce |--+ +--------------+ | (Snowflake) |
+---------------+ | +-------+-------+
+---------------+ | |
| Shipping |--+ v
+---------------+ +--------------+--------------+
| | |
v v v
+---------------+ +-----------+ +---------------+
| Feature | | AI | | Dashboard |
| Store | | Agents | | API |
+-------+-------+ +-----+-----+ | (FastAPI) |
| | +-------+-------+
v v |
+---------------+ +---------------+ v
| Forecast | | Allocation | +----------+
| Ensemble | | Decisions | | Operator |
| (Prophet + | | (Reorder, | | Portal |
| XGBoost) | | Splits) | | (Next.js)|
+-------+-------+ +-------+-------+ +-----+----+
| | |
+--->-------------+ v
Feeds allocation +----------+
| Client |
| Portal |
| (Embed) |
+----------+
The ingestion tier runs on Airflow with dbt handling the transformation layer. Kafka streams updates from the WMS and e-commerce systems in near-real time so the canonical model is at most a few minutes stale. The contract tests are enforced at the dbt layer — a bad source upstream surfaces as a failed build, not as a silently corrupted dashboard.
The forecast ensemble runs on a daily schedule with intraday triggers for high-velocity SKUs. Features live in a dedicated feature store so the training and serving paths stay aligned. The allocation agent consumes forecast and current inventory state and writes reorder and allocation decisions back into Arison's planning workflow — with the exception-handling agent surfacing anything that falls outside confidence bands for a human planner to approve.
The dashboard is a Next.js application backed by a FastAPI service. Every metric visible in the UI is exposed as a stable API endpoint, which is how Arison's larger clients now embed inventory visibility directly into their own procurement portals — no phone call, no email, no CSV export.
04 · Results
Working capital reclaimed, stockouts near-eliminated, clients self-serve
Working capital tied up in inventory fell 35% over the first full quarter of operation — the single largest CFO-visible outcome.
Stock accuracy rose from 89% baseline to 98%, measured against independent cycle counts. The gap closed across every SKU density tier.
Turnover rose from 5.2x to 7.4x. Same revenue base, less capital tied up, faster catalog refresh, healthier margins.
Down from fourteen events per month at baseline to two — and both remaining events were supplier-side failures outside Arison's control.
The dashboard refresh target — under two minutes from warehouse event to visible in the UI — was the detail that unlocked client trust. Clients stopped calling to check stock because the numbers they could see were numbers they could act on. Client NPS from Arison's support function moved sharply positive as clients self-served inventory queries, and support headcount was redirected to higher-value account work.
Internally, the planning team shifted from firefighting to forecasting discipline. The allocation agent handled the routine reorder decisions, which freed planners to focus on the exception cases the agent surfaced — the long-tail SKUs, the promotional bundles, the new product introductions where human judgment actually adds value.
The CFO approved next-phase expansion within 60 days. The working-capital release alone cleared the internal hurdle for the program. Two of Arison's larger clients are now in contract negotiation to embed the client portal directly into their own procurement systems — a commercial outcome that was not even in the original scope.
The KPI-first framing is what convinced our CFO. We had real ROI defended in under 60 days and warehouse visibility we had chased for three years.
Logistics Leadership · Arison
Technologies deployed
- Prophet
- XGBoost
- Airflow
- dbt
- Snowflake
- FastAPI
- React
- Next.js
- AWS
- Postgres
- Kafka
- Grafana
Planning a similar system?
Book a 30-minute strategy call. We will stress-test the use case and share how we would approach it before you commit.